深度学习最新进展.ppt

How does Reduce the Dimensionality of Data with Deep Neural Networks？ * ◆ 2012, ImageNet Recognition Challenge, Hinton’s team with a 16% error rate in classifying 1.2 million images, against a 26% error rate by its closest competitors. ◆ 2006, Hinton and Salakhutdinov proposed deep neural networks learning method; Subsequently (June 26, 2012, news), researchers at Google and Stanford University created the largest neural networks which is a deep neural network with nine-layer that learned on the internet to learn on its own. The system was used to simulate human brain, with one billion connections, and trained over three days on 10 million images by connecting 16,000 cores. After three days’ unsupervised learning, it invented the concept of a cat. (Oct, 2012, news) Microsoft display a fully automated interpretation system on the 21st Century Computing Conference. The Key technology of this system is DNN(Deep neural networks). ◆ 2013, Hinton-Google; Yann LeCun -Facebook; Baidu（IDL） ◆ 2014, Andrew Ng -Baidu Recently Advance of Deep Learning ◆ 2015, Jorunal citation reports for SCI Three Ideas of CNN 1. The area of local perception Fig.1 Connected way of Conventional current neural nets Fig.2 Connected way of current neural nets 2. Share weights Fig.3 Single convolution kernel Reduce the complexity of network and reduce amount of calculation Fig.4 Multiple convolution kernel Fig. 5 Inside a convolutional neural network 3. Secondary extraction of the feature maps Research points that can produce new ideas ※ Initial weights pretraining algorithm ※ The architecture of deep neural networks ※ Layer-by-layer learning algorithm Reference [1]G.E. Hinton, R.R. Salakhutdinov. Reducing the Dimensionality of Data with Neural Networks [J]. Science,2006,313(28):504-507. [2] Yann LeCun, Yoshua Bengio, Geoffrey Hinton. Deep learning [J].Nature, 2015, 521 (28): 436-444. [3] /stdcoutzyx/article/detailsTh